A Physics-Informed Prediction of CO2–Oil Minimum Miscibility Pressure Using Dimensionless Groups and Optimized Neural Architectures
The Minimum Miscibility Pressure (MMP) between injected CO2 and reservoir oil is a critical parameter for the successful design and implementation of miscible CO2enhanced oil recovery (EOR) projects. Accurate and rapid determination of MMP is essential for optimizing injection pressure and maximizing sweep efficiency. Traditional experimental methods, such as the slim-tube test, are time-consuming and expensive, while existing empirical correlations often lack generalizability due to reliance on limited data or simplistic fluid characterization. This study proposes a novel, physics-informed approach for predicting CO2–oil MMP by integrating reservoir fluid thermodynamics with advanced machine learning techniques. Dimensional analysis was first used to identify key dimensionless groups that inherently capture the complex interplay of reservoir temperature, oil composition (characterized by molecular weight and mole fractions of C5–C6 and C7⁺ components), and CO2properties. These dimensionless groups serve as physically meaningful input features, significantly reducing the dimensionality of the problem and improving the physical consistency of the model. Subsequently, optimized Neural Architectures, specifically a Bayesian-Optimized Deep Neural Network (BO-DNN) and a Physics-Informed Neural Network (PINN), are developed and trained on a comprehensive dataset of experimental MMP values. The BO-DNN is optimized for hyperparameter selection to maximize predictive accuracy, while the PINN incorporates the relevant phase behavior constraints and equations of state (EoS) as soft constraints in its loss function, enforcing adherence to fundamental thermodynamic principles. To evaluate performance, the proposed Physics-Informed Neural Network (PINN) and Bayesian-Optimized Deep Neural Network (BO-DNN) were benchmarked against industry-standard empirical models, including the Yellig-Metcalfe, Glaso, and Alston et al. correlations. The results demonstrate a significant paradigm shift in predictive accuracy. While the Alston et al. and Glaso correlations exhibited limited reliability with coefficients of determination (R2) of 0.78 and 0.81 respectively, and Mean Absolute Percentage Errors (MAPE) often exceeding 18-22%, the proposed dimensionless framework achieved an R2 above 0.94. Specifically, the optimized neural architectures achieved a MAPE reduction of 15% to 25% compared to the best-performing traditional algorithms, significantly lowering the Root Mean Square Error (RMSE). The PINN, in particular, demonstrated superior predictive stability by enforcing Equation of State (EoS) constraints, preventing the physically inconsistent "drifting" often seen in legacy formulas. By providing a high-fidelity, cost-effective alternative to laboratory experiments, this research offers a resilient tool for real-time field optimization, aligning with the digital transformation goals of the contemporary energy landscape.